This is a report on the Swiss {ai} Weeks hackathon in Bern, starting with a philosophical reminder of the purpose of a civic hackathon, ending with detailed results and follow-up references. Interspersed with visuals and social media posts, at the bottom you will find links to our complete sources, galleries and hashtags.

.. The F{ai}R
... Presentations
.... Conclusions
..... Gratitude

Apropos hackathon

A few words about the format, a recurring theme of my blog. Why do we hack at a hackathon? Well, we may exlaim, it is because ... Freedom! What kind of freedom, then?

Free from obligations to schools or employers or families? Often we can only participate after negotiating the time-out as an educational or otherwise valuable experience.

Free from having to represent something? People often identify socially, wearing clothing, or having something on their badge, a logo of an employer or a community.

Free from loyalties to institutions? Typical challenges have institutional provenance, being provided by sponsors or academia. Overground hackathons can only function when people are respectful of the hosts and organizers, of the law of the land, of ethical principles and Codes of Conduct.

Free from technical debts? Don't get me started ...

Let us contrast the experience with another widespread practice in our democracy. On this last weekend, a momentous one for digital Switzerland with the narrowly accepted e-ID referendum, I was called in to help count up parliamentary votes in the municipal election. About a hundred of us were told exactly where to sit and what to do. While on duty, nobody was allowed to leave the room. We were instructed with precision in an atmosphere of speedy purpose. Everything one did got checked and cross-checked, in a sturdy ritual of communal validation.

One may suppose that the absence of such control gives one a freer sense of agency? You could think of the civic hackathon as the antithesis of an electoral commission, though: I would argue, in some ways they are complementary. A public process to evaluate technological boundaries, file issues, voice concerns, vote with your attention and time-commitment, should also be seen as a pillar of an advanced society. Consequently, the hackathon setting needs rigor and structure to function – greater than just anonymity, the absence of controls, or precise obligations on their own.

While hackathons and Swiss democracy may seem like polar opposites – one a space of chaotic creativity, the other a model of structured governance – they both underscore a deeper truth about freedom. Freedom isn't about the absence of rules, but about having a framework that allows for agency, innovation, and accountability. In a hackathon, the rules are there to protect the freedom to explore and create, just as in Swiss democracy, the rules ensure that every voice is heard and every vote counted.

Both celebrate the human desire for agency and community, albeit in different contexts. The freedom to innovate in a hackathon complements the freedom to participate and be heard in a democratic society—both are essential for a vibrant, equitable society. The clock ticking towards a deadline being the universal constraint with which hackathons, and democratic elections, make their mark.

As a technologist involved in competitive programming for some 30 years, I would say the most important freedom of a hackathon is that of software freedom, technical choice, the hacker ethic: that I can stick to what I know, or adopt new alternatives. Use well established methods, or walk the path less travelled. At an open community event, I have the freedom to explore and criticise faults openly – within a supportive peer group, rather than as a lone bounty hunter. Other people will (naturally) have other reasons, and that diversity of backgrounds and balance of freedoms and constraints – not just the free drinks, food, space and swag – make the hackathon a unique venue.

In this respect, here is another one marked on my calendar: the 40th anniversary hackathon with the Free Software Foundation (FSF), two months from now at the end of November: "Free software projects and hackers at any stage of their development are invited to participate." .. Nota Bene: "The use of machine learning, like Copilot, ChatGPT and the like, is not allowed." A reminder that not everyone seems freedoms the same way, but that we are free as a society when we exercise our rights, and take part in the debate.

As mentioned in prior blog posts, the Swiss {ai} Weeks are a pioneering effort – similar to past Digitaltage (Digital Days) and other national cyber-expos. This broad alliance formed around the topic of A.I. built for the public good, and on the framework of the Sustainable Development Goals. Nationally, there are over two hundred happenings in enthusiastic response. As Regional Coordinator in Bern, my focus was to organize a trade fair and hackathon last week. Without any further diversion, I will now dive into the results of these events.

Enter the F{ai}r Hackathon. Free from the certitude of a user handbook, or carefully planned out requirements document, we make fresh plans, we bootstrap, copy-paste, vibe code, to Get Things Done in a spontaneous, improvisational manner that seems to me more like a jazz band's jam session - than a crunch session of an engineering team (* we hope that both will learn much from one another)

The F{ai}R at #siliconlovefield

To launch the new hashtag, we printed t-shirts and ran a local trade fair. Space was made for a dozen stands, a day reserved for presentations from representatives. There were several cool workshops: from embedding chatbots into a business model, to coding live agents on a real-time battelship board. We learned how to prompt creatively and effectively in a competitive prompt battle, and more!

Sandro and Selma were amazing hosts for us last week, organizing hot and cold food and drinks at their bar, complemented by Vi’s cuisine in the food truck outside. If you have a chance to come visit Silicon Lovefield, our new hub in Liebefeld, I encourage you to stop at WORKSPACE & MORE - the most light-filled, greenest, quietest and geekiest coworking space in the region.

The F{ai}R started early on Monday, and we reserved Thursday and Friday for the hackathon - two hackdays connected to the presentations and exhibitions. Some people put in ideas, others came back to also take part in the hackathon, energized with their learnings and questions. You can find more impressions in my previous blog post:

Or just watch the SRF coverage of the event to get a 5 minute impression of the venue:

Dribdat = Our Platform

At the start of the Bern Hackathon, we presented 10 challenges across a range of ideas and sectors. Anyone could submit an idea through a series of workshop sessions in the summer. We received over 20 submissions this way, and also encouraged continuing challenges from previous hackdays. Eight of the challenges were finally worked on by the 92 people who registered and created a user profile.

The Dribdat site is where you will find a description of every challenge and resulting project. This is an open source, self hosted web application that we have been developing in support of open data hackathons for years. It was proposed and evaluated as the Swiss-made platform of all the Swiss {ai} Weeks hackathons. Unfortunately, it was not possible to get broad support, so it was only used in Bern. The other hackathon organizers each chose a different platform, as far as I'm aware no other site used a Swiss-hosted solution: something that, I trust, will be discussed in retrospective.

Our Dribdat instance was hosted by Ungleich on a 100% carbon-neutral data center, connected to Hugging Face for authentication, and to the Apertus 70B instruct model provided by PublicAI. In the introductory session, we encouraged people to use their choice of collaboration tool, and announced a total of four alternative options for accessing Apertus: including local, national and international providers of AI services. All these, and 25 other technical platforms that were supported at the hackathon, are listed in the Resource wiki. You can read all challenge proposals, including the ones that did not fit into this year's hackathon, on Dribdat.

Habemus Apertus

Complementing the launch events from the Swiss AI Initiative, we had a challenge that was completely Apertus-generated, based on the description of our hackathon. It was spoken out by an AI voice from ElevenLabs, on stage (representatively) a small Otto DIY open-source robot. That was quite a nerdy moment at the end of the presentations, even if a team didn't form around the idea. Perhaps next year, when we get a more expressive robot as your coach and fun mechanical team-mate, this type of AI challenge will have a chance 🤖

It was interesting to see this auto-generated proposal to "consider as a team how to truly meet the needs of global communities by enhancing specific cultural capabilities - accessing untapped datasets, or even advocating for data contributions like it was done already for Rumantsch dialects". It reflected the vision manifest (Appendix O) of the Apertus Technical Report, as well as the system prompts of the PublicAI deployment.

While several of the mentioned data sources were hallucinated (m4cite ? Erasmus+ ?), the contact information at the end reads like an Easter egg from the designers, inviting hackers to connect with improvement suggestions – the hackathon providing an excellent venue to draw attention and new recruits.

Indeed, the Habemus Apertus challenge was something that you could say was in the background of all our projects. The new Swiss-made LLM model is a big deal: you can read all about it in my earlier blog post, including instructions of how to get it running on different services and your hardware. We were fortunate to have the expert guidance in this of Prof. Marcel Gygli (presentation, slides) in a tech session on Thursday.

There was no requirement to use Apertus at the hackathon, but we can assume that everyone tried the free chat. Several of the teams explored model capabilities and used it seriously, and there were several instances of proposed improvements. Questions about the reliability of results, provenance of the datasets, hosting and training options for the Apertus model came up. We had a physical copy of the Tech Report on site, a Hugging Face forum, GitHub repo and Discord server where we could forward tricky questions to the experts.

Prompt Buskers

This was a 'meta' challenge where we aimed to collect and express creative ideas during the event. Note that it was generated by Mistral 24B and inspired by Buskers Bern, which took place a few weeks before we had access to Apertus.

We invited a local musician - CyberGwen, who first played at the F{ai}R, then performed for nearly two hours on Thursday evening. By thinking and planning ahead of time to involve people from diverse backgrounds, bringing in a cultural performance at a critical juncture, it was a small but meaningful contribution. I hope that it will inspire you to organize creative workshops, and support local entertainment artists at future hackathons.

Watch CyberGwen tune up the night with open source visualizations here - see full credits in the video description:

Team Presentations

Our participants were asked to split into groups of maximum 5 people, though it was more of a rule of thumb rather than strictly enforced. You can see the directory of 92 people who created a profile on Dribdat. The following 9 project teams presented their results, in this order, on Friday:

1. Energy Infrastructure from Remote Sensing Team Beta
2. Guardrails as Code
3. AI Mates powered by Apertus
4. Archive Image Matching 🏆
5. Create your own Planetary Systems
6. Local produce transportation 🏆
7. Measure footprint of open LLMs
8. Tibetan Chatbot 🏆
9. Energy Infrastructure from Remote Sensing Team Alpha

In the following sections, I review each project in a bit more detail. You can also skip the first 13 minutes to watch the presentations in our video recording of the presentations, or watch my 20 minute short video review of the documentation.

Energy Infrastructure from Remote Sensing

From this year, residents of the Canton of Bern are not obligated to register their solar panels, and two teams (Alpha and Beta) worked on the problem of getting better insights into energy production. The goal of the challenge, proposed by the environment and energy commission, is to use open data to estimate production capacity, and create tools to help with energy planning and communications.

We had a chance here to learn a lot about the topic of collecting urban indicators, and exploring them with an A.I. system. They first wanted us to work with sophisticated, highly granular datasets, but ended up simplifying things a little bit. There was actually a whole hackathon run one week before ours on the topic of energy data, also part of this year's Swiss {ai} Weeks: the Energy Data Hackdays. I was there and coached one of the teams.

One gets the sense that this is a huge topic: how do we address the climate change crisis with more resilient infrastructure in terms of energy generation and consumption? You can see a little bit in the log that this team struggled to have the right computing capacity. Both teams worked really hard on on the challenge, training a model with a new AI architecture to improve the rate of detection of solar panels, then plotting the results on a map to see how effectively the cantonal policies are being implemented in various regions.

The Alpha team collected and labelled images of roofs in Bern, aiming to train a data model that consistently recognized patterns in energy production facilities. In the Beta team's presentation you can find a Readme (in Italian), which describes how they acquired data from sources like Swisstopo, extracted it and analyzed the orthophotos. Both teams used computer vision techniques - similar to the Archive Image Matching team - to detect the energy infrastructure from satellite pictures. Very interesting research attempts were made, along with supporting visualizations and prototype dashboards.

Team Alpha

Recording from 1:14:00

Team Beta

Recording from 14:00

Guardrails as Code

Technical guardrails are safety measures designed to ensure that artificial intelligence systems operate ethically, safely, and within defined boundaries: functions and practices to help certify that a service is compliant. For example, to help ensure your chatbot or automation stays within more or less strict delineations of what it is supposed to know, chat, or decide about. Interprimis, a local consultancy, challenged us to apply them to AI.

In the case of Swiss and European laws (think A.I. Act) this may be data protection issues, for Internet content you may define a scale of toxicity ratings. The project addresses to various issues they see in implementing guardrails. The team conducted some research, generated a proposal with OpenAI which discusses the impact and costs of implementing - or, indeed, the risks of deactivating - guardrails in a design blueprint. I would have liked to see a small demo, but they lacked a developer in their team. They shared resources, clarified requirements, and expressed a readiness to involve people in the future in addressing a critical topic that is a part of most LLM systems right now.

Recording from 21:30

AI Mates powered by Apertus

A dating platform idea with a twist! Not just another Tinder clone: here you have an authentic challenge, an audacious idea, a true desire to help people and apply personal experience. The result was arguably the most dramatic and memorable presentation on Friday, with the team acting out their arguments to revolutionize modern romance. The Agentic approach of the solution is the hottest thing in the business right now: in this app, you will create an avatar, sharing your data with a virtual agent that goes out and "meets" others like it, sharing learning about the experiences before you go out on a real date.

The team worked intently on a prototype using the PublicAI and Brandbot instances of Apertus, and came up with a convincing click-prototype. They thought a lot about what it would take to implement it, worked on extensive product requirements. My time as organizer with any of the individual teams was very limited, but in the night during the hackathon I vibe coded a mock MCP (model context protocol) server. We will see soon if this is one of the approaches we can take. On the whole, a nice initiative, clearly something that the whole team is passionate about, and sees a market ready for. Let's AI Mate!

Recording from 28:40

Archive Image Matching

A project from the University of Bern, that was done by a team of people who wanted help to categorize images in a very large archive of hundreds of thousands of pages of historical prints. In the past, this would have been a popular crowdsourcing task, today the goal is to distribute the work among AI tools. There are many issues and constraints to work with to get these texts read in properly with advanced computer vision, and clustered with LLMs.

The solution was to augment traditional techniques with new algorithms to optimize term frequency and improve the accuracy of the results that they're getting out. They trained a very high performance AI model that takes half a second to process an image. To correctly manage the matched images, they also created a frontend and backend catalog. This is a FastAPI based server connected to Apertus, with a bit of code that allows searching, uploading, and better understanding the content - and lots of cool little features built in. I really appreciated seeing a working demo with Brandbot running Apertus. Very cool that they came up with this, and a big value provided to the university team.

Recording from 37:00

🏆 Congratulations to the team for winning the Public Vote!

Create your own Planetary Systems

This team worked on an AI algorithm developing an understanding of the structure of exoplanetary systems - planets like our own in other solar systems. In a fun set of presentation slides, starting with this ancestral person looking up at the stars wondering is there life out there? You’ve got to love space science. Lots of data, different telescope arrays and "are we alone?", the deep and searching questions that robots should help us with. We were a rapt audience!

The project was discussed again after our hackathon in the on{ai}r webinar. In the video linked above, they talk about the general topic of exoplanet research, how crowdsourcing helps the discovery of other Earth-like planets, the role of AI, and review the project in depth. Multiple prototypes came out of this this team, both an intense statistical model with deep number crunching and data analysis, as well as an interactive solar system you can play with. Don't miss their Dribdat log to see all kinds of outputs from their hackathon experience that did not make it into the final presentation.

Recording from 43:30

Local produce transportation

"Why don't you order your food from a local farm shop?" Starting with this simple question, the team worked to create tools for shops and farmers to easily transport their goods to customers. In their presentation they presented Farmly - even coming up with a logo and brand – to mock up an app. It allows chatting with an intelligent AI service, that accepts tasks and cues them into an ordering platform, where you can order various fruits and vegetables or farm products. The route that these products would take is calculated using the OpenRouteService API.

They built and demoed their solution - mainly designed for farmers, thinking out in lots of detail how it would work. Great results from a team of young people who worked without their challenge owner, as that person unfortunately got sick just before the hackathon. Very courageous of them: and doubly so to make a national television appearance with an in-depth interview in the middle of the hackathon. Hope to see a launch page for the product up and running soon!

Recording from 50:00

🏆 Congratulations to the team for tying for top in the A.I. ranking!

Measure footprint of open LLMs

This team set itself as a goal to really understand how Apertus uses energy, and got under the hood to benchmark the LLM, compare it with other models, and recommend strategies for efficient prompting. The energy profile of AI as being one of the top concerns people have in using the very energy-hungry services that incorporate it, it was very interesting to hear what the team had to say about energy consumption – and to what extent we as everyday users can influence it.

We had various infrastructure made available: a large Mac Studio running in a data center was sponsored by Begasoft. There were also two graphics workstations on site, which we had trouble getting to work. The model is still quite new and doesn't easily run on all hardware architectures: see my separate post for more detail.

The experiments that the team ran were to test different prompt and response lengths, different subjects, a number of different languages, etc. They ran comparisons with Llama, another popular open source model, to try to really understand how much energy is being used in different use contexts. Unfortunately, they were not able to get a very stable setup. For a good benchmarking, you have to be able to reproduce the results over and over again. They were very forthcoming with their failures: the basis of future efforts.

The team figured out how to measure a range of different consumption values through local and remote access. They clearly learned lots, had fun, created graphs, and shared everything in their project page – which is all great to see at a hackathon. Some very good inputs overall from a team that took a difficult subject, tackled it, and ended up with something that we can learn from and use going forward.

Recording from 58:20

Tibetan Chatbot 

Again focusing on the use of the new Swiss LLM, which states as one of it's main features a strong linguistic ability, the challenge here was to build a RAG with new multilingual references. This team really applied itself to feed in new data from the Tibetan language, working with Unicode text that they got out of a large spreadsheet (open data available on GitHub), or other resources found online. They prepared this to create a chatbot for language learners.

The team used the Apertus 8B local model, finding that it works no less well than the 70B for the specific requirements of their app. They tested various configurations including local deployment, showing the edge computing potential of more lightweight (and less energy-intensive!) AI models. Their demo using Streamlit was tested with excellent effect live in the hackathon presentation. A screencast will be available on their project page.

The idea of having an AI tutor is one we all understood, and it was to see a live demo that we could play with during the hackathon. The Tibetan learning community will hopefully get to benefit from these additional tools. The fact that Apertus is in the center of attention here is a great call to action, seeing whether the model delivers what it promised. I thought this is a promising project, that we are going to be hearing more about soon.

Recording from 1:06:00

🏆 Congratulations to the team for tying for top in the A.I. ranking!

Conclusions

It's hard not to be impressed with the energy and dedication of every single team, yet for various reasons ... this is what I would expect from a well supported hackathon. From the moment the doors opened until the final project pitches closed, every member of every team had the opportunity to get fully engaged, sharing ideas and working together with remarkable camaraderie. We had nine brilliant presentations, during which you could see evidence of a truly cooperative setting, encouraging everyone to pitch in regardless of technical proficiency.

What impresses me the most was the community spirit—whether it was teams solving complex AI challenges side by side, learning from each other, or supporting one another's ideas. A cooperative energy that keeps us organizers going, and inspires future events. The excellent location and catering we had is available for this, so get in touch if you'd like to run your own hackathon in Liebefeld!

To add some depth, we assessed all the projects using an AI-driven evaluation based on the five criteria of the Swiss {ai} Weeks hackathons: Technical Functionality, User Experience, Skillful use of AI, Uniqueness / Creativity / Fun Factor, and Potential / market impact — all scored autonomously by Apertus 70B. These evaluations were based purely on the project documentation, offering an impartial complement to the public votes. If you are interested in AI evaluation of ideas and presentations at your event, let us know.

The event was covered by SRF, Switzerland’s national broadcaster, and we uploaded a photo gallery to give you a visual taste of the hackathon atmosphere. You can find the Apertus reviews in the Log (Dribs) of every project. These three projects earned special recognition, and the team members earned small prizes:

Top in Public voting:

• Archive Image Matching: A project that leveraged AI to help sort and organize historic images, enhancing accessibility and search capabilities within archives.

Top in AI evaluation (tie):

• Local Produce Transportation: A chatbot that aimed to help local farmers find convenient transportation solutions using AI-powered matching systems.
• Tibetan Chatbot: Developing an AI tutor based on the Apertus model to support access to content and learning of the tibetan language.

To learn more, click through the project descriptions and presentations on our event page to see the full scope of innovation. Discover the 27 guides we assembled in the Resources section with explanations of accessing tools from Apertus and Supertext, detailing how they could be used during the hackathon. Feel free to reuse everything under Creative Commons license, for your own event.

Gratitude

Thank you for supporting us - whether you joined us at the F{ai}R & Hackathon in #SiliconLovefield, were part of another event in Bern or another of the 24 cities represented, cheered us on from the social network sidelines, or even stopped by to read this blog post and other coverage! The Swiss {ai} Weeks were a massive undertaking, and it will take time to fully process all the impact.

Special thanks to Selma & Sandro for being our brilliant hosts at WORKSPACE & MORE. We are all grateful to Prof. Dr. Marcel Gygli (BFH) for the subject matter workshop, and to CyberGwen (YouTube) for an incredible concert on Thursday. To the Mê food truck for delicious vietnamese cuisine, and Harry Stitzel (SRF) for intrepid reporting on Friday. To Kim Chai Ly, Jürg Stuker and Pascal Mercel - high five's for critical support at key moments. To all of you who put in your time, energy, human openness and intelligence: your contributions will be remembered.

The event was made possible through volunteers backed by the Economic Development Agency of the Canton of Bern, with additional financial contributions from fers stiftung and Puzzle ITC. Technically, our AI hackathon was extremely well supported by the Swiss AI Initiative, Begasoft, Swisscom, Elevenlabs, Supertext, Hugging Face, Init7, Ungleich / Dribdat and PublicAI.

There are people in the background without whom none of this would have happened - or at least not for a long time: big shout-outs to Sabine Wildemann, Diana Engetschwiler, Daniel Dobos, Christoph Birkholz and team, investing massive efforts to create the Swiss {ai} Weeks, bringing people together who want to shape the future together.

It was a momentous week of exchanges, inspirations, a possibility to test fresh ideas hands-on, and meet people from Bern, all over Switzerland, and around the world. If you are interested in more events like this, use the Hackfinder at hackintegration.ch

Links to recordings and reports

Please get in touch if there's anything I've missed!

• Team Presentations
• Pitch Workshop - SLIDES
• Expert Workshop - SLIDES
• Opening & Challenges - SLIDES
• SRF 10vor10 Report (in German)
• Photo gallery - Swiss {ai} Weeks
• Dribdat ⬡⬢ - our hackathon platform
• LinkedIn - #swissaiweeks #siliconlovefield
• Instagram - #swissaiweeks #siliconlovefield
• Mastodon - #swissaiweeks #siliconlovefield

Once more: thank you to everyone involved, from organizers, participants, and judges to the university and media partners. Whether you were there or watching from afar, I hope you found the Swiss {ai} Weeks scene to be vibrant, open, innovative, and full of promise. We are all excited for what is coming in the years ahead.

For now, dive into the open projects, share your thoughts. Let’s continue the conversation on our social channels or here in the comments. Stay curious, stay cautious, stay freedom-loving, stay creative .. let’s keep building the future of AI in Switzerland together!

[Apertus 70B was used to support the composition of this blog post]